Scholarlink

To address the security challenges in in-vehicle networks, this paper proposes a lightweight intrusion detection method that integrates MobileNetV3 and GRU, further enhanced by a hierarchical knowledge distillation mechanism to improve feature representation. In the data preprocessing stage, a multi-channel image encoding strategy is employed to convert raw CAN messages into structured four-channel images, effectively capturing semantic, temporal, and abrupt anomaly information. The proposed model leverages MobileNetV3 to efficiently extract spatial features, while GRU captures temporal dependencies within message sequences. Additionally, the hierarchical distillation module enhances the model’s representational capacity and discriminative power without compromising its compactness. Experimental results on the Car-Hacking dataset demonstrate that the proposed approach outperforms mainstream models—including CNN, LSTM, RF, and EfficientNetB0—in terms of accuracy (98.86%), F1-score (98.65%), and inference efficiency, indicating strong detection performance and practical deployment potential.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lightweight Image-Based Intrusion Detection for In-Vehicle Networks Via MobileNetV3-GRU and Hierarchical Distillation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zhihui Cai,
  • Guangfu Wu,
  • Changqin Xu

摘要

To address the security challenges in in-vehicle networks, this paper proposes a lightweight intrusion detection method that integrates MobileNetV3 and GRU, further enhanced by a hierarchical knowledge distillation mechanism to improve feature representation. In the data preprocessing stage, a multi-channel image encoding strategy is employed to convert raw CAN messages into structured four-channel images, effectively capturing semantic, temporal, and abrupt anomaly information. The proposed model leverages MobileNetV3 to efficiently extract spatial features, while GRU captures temporal dependencies within message sequences. Additionally, the hierarchical distillation module enhances the model’s representational capacity and discriminative power without compromising its compactness. Experimental results on the Car-Hacking dataset demonstrate that the proposed approach outperforms mainstream models—including CNN, LSTM, RF, and EfficientNetB0—in terms of accuracy (98.86%), F1-score (98.65%), and inference efficiency, indicating strong detection performance and practical deployment potential.